Martin Stommel

Binarising SIFT-Descriptors to Reduce the Curse of Dimensionality in Histogram-Based Object Recognition

It is shown that distance computations between SIFT-descriptors using the Euclidean distance suffer from the curse of dimensionality. The search for exact matches is less affected than the generalisation of image patterns, e.g. by clustering methods. Experimental results indicate that for the case of generalisation, the Hamming distance on binarised SIFT-descriptors is a much better choice.

A Novel Soft Machine Table for Manipulation of Delicate Objects Inspired by Caterpillar Locomotion

This paper presents a soft XY machine table inspired by the locomotion of the caterpillar. The proposed table is capable of manipulating objects in three degrees of freedom through the deformation on the soft surface. The softness of the surface enables the table to handle delicate objects such as food products and industry components without damaging them. A novel object manipulation method is inspired by the proleg movement seen in the caterpillars crawling gait. A prototype of the tables actuation system has been designed and fabricated. Our experimental results prove that the table is able to move an object in both translational and rotational direction. This validates the manipulation concept developed and shows that the challenging task of transporting flat objects is accomplished. A representation method for surface deformation is developed to aid the modeling and control of the physical system in the future.

Inpainting of Missing Values in the Kinect Sensor's Depth Maps Based on Background Estimates

Low-cost game controllers such as Microsofts Kinect sensor provide dense, real-time depth measurements of indoor environments at high framerate. The sensor is based on the principle of active stereo using structured infrared light. For surfaces that distract infrared light, no measurements can be obtained. It is important to replace missing values early on, to avoid a slow subsequent conditional evaluations or the propagation of errors into neighboring regions. To solve this problem we present an inpainting method that adds missing values based on background estimates of the unoccluded scene. It is therefore not necessary to hypothesize missing regions based on similarity to other image regions. The procedure also avoids a blurring between foreground and background. By adapting the method to the specific properties of the Kinect (and comparable) cameras, we were able to keep the complexity of the algorithm low, so high speed can be achieved.

Optimal, Efficient Sequential Control of a Soft-Bodied, Peristaltic Sorting Table

A peristaltic, soft-bodied xy-sorting table manipulates objects by producing moving wave shapes on its surface. The waves exert forces on the objects which can be used for transportation, reorientation, and local repositioning. The control of such peristaltic robots is mostly unsolved, because important properties of the kinematics, dynamics, and effect of actuation are unknown. Fundamental and practical limitations in measuring the system state lead to numerical difficulties in the form of discontinuous signals in non-Euclidean spaces. To solve these problems, we introduce a probabilistic automaton that models the static and dynamic effect of actuation based on a discretized representation of the objects to be manipulated. The automaton only considers the qualitative input-output behavior of the system. It is therefore mostly independent of a particular hardware setup and controls the behavior of objects on the table without explicitly predicting mechanical forces. Our theoretical findings advance the field significantly: We show that the proposed class of automata is suited to model both static and dynamic movements. We introduce a cost function that enables the search and identification of optimal sequences of control patterns to bring the system into a desired state. We prove that the optimal control sequence can be found efficiently and give the respective algorithm.

Segmentation-free detection of comic panels

The detection of comic panels is a crucial funcionality in assistance systems for iconotextual media analysis. Most systems use recursive cuts based on image projections or background segmentation to find comic panels. Usually this limits the applicability to comics with white background and free space between the panels. In this paper, we introduce a set of new features that allow for a detection of panels by their outline instead of the separating space. Our method is therefore more tolerant against structured backgrounds.

Model-Free Detection, Encoding, Retrieval, and Visualization of Human Poses From Kinect Data

The recognition of humans in Kinect camera data is a crucial problem in many mechatronics applications with human-computer interaction. In order to improve the limited scope of many methods based on a kinematic or surface mesh model, we propose a spatiotemporal segmentation of keypoints provided by a skeletonization of depth contours. A vector-shaped pose descriptor allows for the retrieval of similar poses and is easier to use with many machine learning libraries. A visualization method based on the Hilbert curve provides valuable insight in the detected poses. Our experimental results show that the proposed method is able to adapt to the number of people in a kitchen scenario, and track them over time. We were able to retrieve similar poses from a database and identify clusters in the dataset. By applying our method, the Princeton Tracking Benchmark, we demonstrated that our method is applicable in scenes where a human kinematic or surface mesh model would be overly restrictive.

Learning of face components in coherent and disturbed constellations

A face recognition system for simultaneous detection and pose estimation is presented. The algorithm proceeds in two steps: At first, separate face components such as eyes, nose and mouth are detected. This is done by a classification of modified SIFT features that are more robust to spatial displacements. Secondly, face-like part constellations are detected by an SVM based voting scheme. Inhibitive votings are introduced to suppress false detections in textured image regions. Experiments on the Feret and Graz data bases demonstrate the high accuracy of the system.

The bitvector machine: a fast and robust machine learning algorithm for non-linear problems

In this paper we present and evaluate a simple but effective machine learning algorithm that we call Bitvector Machine: Feature vectors are partitioned along component-wise quantiles and converted into bitvectors that are learned. It is shown that the method is efficient in both training and classification. The effectiveness of the method is analysed theoretically for best and worst-case scenarios. Experiments on high-dimensional synthetic and real world data show a huge speed boost compared to Support Vector Machines with RBF kernel. By tabulating kernel functions, computing medians in linear-time, and exploiting modern processor technology for advanced bitvector operations, we achieve a speed-up of 32 for classification and 48 for kernel evaluation compared to the popular LIBSVM. Although the method does not generally outperform a SVM with RBF kernel it achieves a high classification accuracy and has qualitative advantages over the linear classifier.

Medically-inspired approaches for the analysis of soft-robotic motion control

Soft-robotic structures and their materials are typically chosen according to a biological example. Medical imaging has been used to obtain 3D models of biological structures to create moulds for production of artificial, soft robotic counterparts. However, it is not enough to simply copy the geometry of these organisms; robots must be able to be modeled, and controlled, such that they can perform meaningful tasks. This involves investigating the robots capability after it has been manufactured. The similarities between the biological and artificial robotic materials allow us to use methods from medical imaging in soft robotics. This paper proposes the use of medical imaging and alternative medical investigation methods for the static and dynamic characterization of soft robots and involves two soft-robotic case studies: a peristaltic pump (swallowing robot), and a peristaltic table. Articulography and manometry are shown to be useful techniques for investigation of the peristaltic pumping robot, and visual 3D scanning is demonstrated for the peristaltic table. Alternative medical investigation methods such as magnetic resonance imaging, computed tomography, and ultrasound are considered as other possibilities that require further investigation.

Technical requirements and conceptualization of a soft pneumatic actuator inspired by human gastric motility

This paper presents the technical requirements and conceptualization of the actuator inspired by human gastric motility. Key features of the stomach motion are explored and described in the viewpoint of engineering. A soft robotic model of the stomach is conceptualized for the purpose of in vitro simulation of the contractile motion of the stomach. Soft lithography and lost wax methods are implemented in the fabrication. Our experiments show that the proposed robot is able to reproduce the required contractile movement.